emorf1003/bam-vl-qwen35-27b

bam-vl-qwen35-27b

LoRA fine-tune of Qwen/Qwen3.5-27B for video question answering, exported from a BAM-VL training checkpoint (step_2023, global_step 2023).

Contents

• —Merged backbone in the standard HF layout — loads with AutoModelForImageTextToText.
• —bam_adapters.pt — the BAM side-channel adapters (facial / pose / audio) and CLS heads, with the hyperparameters needed to rebuild them.

Important

The merged backbone alone does not reproduce training behaviour. During training the model pools the penultimate hidden state, computes a delta from the side-channel adapters, and adds it to the final hidden state before lm_head. A plain generate() call skips this entirely. Faithful inference requires bam_adapters.pt together with the BAMVLQA wrapper code and the corresponding facial/pose/audio feature extractors.

Training

• —Stage 1: bam_only — frozen backbone, side-channel adapters only.
• —Stage 2: bam_and_full_model — LoRA (r=32, alpha=64) plus adapters.